PhD Final Defense – Charbel Beaino

Development of Empirical, Numerical, and Machine Learning Approaches for Estimating Deep Excavation Response

Advisor: Professor Youssef M. A. Hashash

Meeting ID: 239 695 320 029 868 

Passcode: iF7YG26Y

Abstract

One of the challenges affecting urban areas is constructing and developing underground space while limiting its impact on the existing structures and utilities in heavily built environments. Excavation-induced ground movement is thus a main concern for such projects, where allowable surface deformations and differential movements are very small. Data collected from seven station excavations for the Los Angeles Metro, and data from the Transbay Transit Center excavation in San Francisco offers a unique opportunity to evaluate excavation support performance. 

When relatively stiff soils and/or excavation support systems successfully control sources of surface settlements, upward surface displacement or heave is observed around the excavations. In soft soils with more flexible support systems, this elastic behavior is typically masked by relatively large settlements due to ground loss and wall deflections. Since surface heave is commonly disregarded, existing empirical methods tend to overestimate excavation-induced settlements and fail to capture the induced heave. A semi-empirical method is proposed for predicting maximum wall deflections and surface displacements of the LA Metro excavations in stiff soils, with shear wave velocity measurements generally ranging between 240 and 460 m/s. By idealizing the excavation as an upward load at the ground surface, the elastic Boussinesq solution is used to provide an upper limit on the estimate of the upward surface displacement. The calculated heave is added to a conventional settlement component. This settlement component resulting from lateral wall deflections is correlated empirically to the system stiffness and the shear wave velocity of the retained soil. The proposed framework enables the estimation of ground surface heave in scenarios where stiff support systems are used to control lateral deflections. This approach is particularly effective in stiffer soil conditions, where traditional methods may not fully capture heave behavior. 

Two- and three-dimensional numerical models are also developed for select excavations. The hardening soil model with small-strain stiffness is calibrated using shear wave velocity measurements and modulus reduction curves to represent the non-linear soil behavior at small strains. The numerical models using the proposed calibration framework are able to represent excavation performance. The three-dimensional model also demonstrates the stiffening and displacement reduction effects of the excavation corners and wall offsets. Simulated and measured movements compare favorably.

In braced excavations, steel struts are simultaneously subjected to lateral earth and water pressures as well as temperature induced loads. Strain gauge measurements on struts reflect the sum of these components, but the individual load components remain unknown. The separation of earth and thermal loads is therefore necessary for quantifying thermal effects. Advances in field monitoring instrumentation have enabled high frequency measurements, offering new insight into strut behavior, but also introducing new challenges in data interpretation. This study presents a framework for evaluating temperature-induced loads using a machine learning-based approach. The proposed approach uses Recurrent Neural Networks (RNN), capable of capturing temporal patterns in the data. The results demonstrate that RNN models provide more stable interpretations of loads compared to conventional linear regression techniques. The results also demonstrate that installation temperature has a direct impact on temperature-induced loads, where low installation temperatures can lead to a large positive thermal change, causing thermal loads to reach up to 18 to 40% of the total strut load.